【作者】 程伟；

【导师】 蒋保臣；

【作者基本信息】 山东大学 ， 电子与通信工程（专业学位）， 2019， 硕士

【摘要】 由于微电网光伏发电功率的间歇性与随机性和用电负荷的波动性会对微电网系统的稳定运行造成较大影响,使得微电网发电端与用电端之间的能量交换变得异常复杂。为保障微电网的安全稳定运行与用户的正常生产生活,对光伏发电功率及负荷进行精确可靠的短期预测具有重要的现实意义。在微电网光伏发电短期预测方面,本文设计了基于相似日优化的长短期记忆(Long-Short Term Memory,LSTM)神经网络光伏发电短期预测模型。首先,对微电网光伏发电特性进行了有效分析并对其影响因素作了充分论证。其次,为提高不同天气类型下(晴、阴、雨天)的光伏发电功率的预测精度,本文基于K-Means聚类算法设计了相似日筛选模型。最后,实现了基于LSTM神经网络的微电网光伏发电短期预测模型,对未来24、48、72小时不同天气类型下的光伏发电功率进行了多尺度、高分辨率的短期预测。在微电网负荷短期预测方面,本文设计了基于相似类型日优化的LSTM神经网络负荷短期预测模型。首先,对微电网典型日、周负荷曲线特性进行了分析并对影响用电负荷的气象因素、日类型进行了量化。其次,设计了相似类型日筛选模型,将初始数据集划分为工作日训练集与休息日训练集。最后,实现了基于LSTM神经网络的微电网负荷短期预测模型,对未来24、48、72小时的工作日与休息日的用电负荷进行了较为全面的预测。通过国标规定的评价指标对上述预测结果进行对比分析,本文所述预测方法的预测结果均满足国标要求。本文光伏发电短期预测结果的平均绝对误差(Mean Absolute Error,MAE)达到了 7.14%,负荷短期预测结果的绝对百分比误差(Mean Absolute Percentage Error,MAPE)达到了 1.75%,相较研究领域内常用的误差逆传播(Back Propagation,BP)神经网络算法的预测精度分别提高了 6.45%与1.12%。本文所述预测方法对微电网光伏发电及负荷短期预测具有重要的应用价值。更多还原

【Abstract】 Since the volatility and randomness of the photovoltaic power generation and the power load of the micro-grid will have a great impact on the stable operation of the micro-grid system,the energy exchange between the power generation unit and the load in the micro-grid becomes extremely complicated.In order to ensure the safe and stable operation of the micro-grid and the normal production and life of the users,accurate and reliable short-term prediction of photovoltaic power and load has im portant practical significance.In the short-term prediction of micro-grid photovoltaic power generation,this paper designs a short-term prediction model of Long-Short Term Memory(LSTM)neural network based on similar day optimization.Firstly,the micro-grid photovoltaic power generation characteristics are analyzed effectively and their influencing factors are fully demonstrated.Secondly,in order to improve the prediction accuracy of photovoltaic power generation under different weather types(fine,cloudy and rainy days),a similar daily screening model based on K-Means clustering algorithm is designed.Finally,a short-term prediction model of micro-grid photovoltaic power generation based on LSTM neural network was designed and implemented.The multi-scale and high-resolution short-term prediction of photovoltaic power generation in different weather types in the next 24,48 and 72 hours was designed.In the short-term prediction of micro-grid load,this paper designs a short-term prediction model of LSTM neural network load based on similar type of daily optimization.Firstly,the typical daily and weekly load curve characteristics of the micro-grid are analyzed,the meteorological factors and day types affecting the power load are quantified.Secondly,a similar type daily screening model is designed,and the initial data set is divided into a working day training set and a rest day training set.Finally,the short-term prediction model of micro-grid load based on LSTM neural network is designed and implemented.and the power load of working days and rest days in the next 24,48,and 72 hours is comprehensively predicted.Through the comparative analysis of the above-mentioned prediction results through the evaluation indicators stipulated by the national standards,the prediction results of the prediction methods described in this paper meet the requirements of the national standard.The Mean Absolute Error(MAE)of the short-term prediction results of photovoltaic power generation reached 7.14%,and the Mean Absolute Percentage Error(MAPE)of the short-term load forecast reached 1.75%.Compared with the Back Propagation(BP)neural network algorithm commonly used in the research field,the algorithm is improved by 6.45%and 1.12%,respectively.The prediction method described in this paper has important application value for micro-grid photovoltaic power generation and short-term load forecasting.更多还原